Meta Open-Sources Muse Glimmer, a 30B Local Agent Model Running on 24GB VRAM
Meta has open-sourced Muse Glimmer, a 30-billion-parameter multimodal AI model optimized for local agent workflows under the Apache 2.0 license. Using 4-bit quantization, the model can run locally on consumer-grade hardware with 24GB or 32GB of VRAM, such as Macs and PCs. This release enables developers to run highly capable, agentic AI workflows entirely locally on consumer hardware without relying on cloud APIs. It lowers the barrier to deploying private, always-on AI agents that can handle complex tool calling and multi-step reasoning. Muse Glimmer features a multimodal architecture supporting interleaved text and images, and it ships with a lightweight DFlash-based drafter for speculative decoding to accelerate inference. It is specifically trained for agentic tasks, including multi-step reasoning, precise function calling, and self-recovery from tool failures.
## BACKGROUND
AI agents are autonomous systems designed to perform multi-step tasks, use external tools, and make decisions to achieve specific goals. Running large models (like 30B parameters) locally typically requires significant memory, but quantization techniques compress model weights (e.g., to 4-bit) to drastically reduce VRAM requirements with minimal loss in performance.